Applications and limits of convex optimization

Every algorithmic learning problem becomes vastly more tractable when reduced to a convex program, yet few can be simplified this way. At the heart of this thesis are two hard problems with unexpected convex reformulations. The Paulsen problem, a longstanding open problem in operator theory, was rec...

Full description

Bibliographic Details
Main Author: Hamilton, Linus
Other Authors: Moitra, Ankur
Format: Thesis
Published: Massachusetts Institute of Technology 2022
Online Access:https://hdl.handle.net/1721.1/145023
_version_ 1826217363837550592
author Hamilton, Linus
author2 Moitra, Ankur
author_facet Moitra, Ankur
Hamilton, Linus
author_sort Hamilton, Linus
collection MIT
description Every algorithmic learning problem becomes vastly more tractable when reduced to a convex program, yet few can be simplified this way. At the heart of this thesis are two hard problems with unexpected convex reformulations. The Paulsen problem, a longstanding open problem in operator theory, was recently resolved by Kwok et al [40]. We use a convex program due to Barthe to present a dramatically simpler proof with an accompanying efficient algorithm that also achieves a better bound. Next, we examine the related operator scaling problem, whose fastest known algorithm uses convex optimization in non-Euclidean space. We expose a fundamental obstruction to such techniques by proving that, under realistic noise conditions, hyperbolic space admits no analogue of Nesterov’s accelerated gradient descent. Finally, we generalize Bresler’s structure learning algorithm from Ising models to arbitrary graphical models. We compare our results to a recent convex programming reformulation of the same problem. Notably, in variants of the problem where one only receives partial samples, our combinatorial algorithm is almost unaffected, whereas the convex approach fails to get off the ground.
first_indexed 2024-09-23T17:02:29Z
format Thesis
id mit-1721.1/145023
institution Massachusetts Institute of Technology
last_indexed 2024-09-23T17:02:29Z
publishDate 2022
publisher Massachusetts Institute of Technology
record_format dspace
spelling mit-1721.1/1450232022-08-30T03:01:24Z Applications and limits of convex optimization Hamilton, Linus Moitra, Ankur Massachusetts Institute of Technology. Department of Mathematics Every algorithmic learning problem becomes vastly more tractable when reduced to a convex program, yet few can be simplified this way. At the heart of this thesis are two hard problems with unexpected convex reformulations. The Paulsen problem, a longstanding open problem in operator theory, was recently resolved by Kwok et al [40]. We use a convex program due to Barthe to present a dramatically simpler proof with an accompanying efficient algorithm that also achieves a better bound. Next, we examine the related operator scaling problem, whose fastest known algorithm uses convex optimization in non-Euclidean space. We expose a fundamental obstruction to such techniques by proving that, under realistic noise conditions, hyperbolic space admits no analogue of Nesterov’s accelerated gradient descent. Finally, we generalize Bresler’s structure learning algorithm from Ising models to arbitrary graphical models. We compare our results to a recent convex programming reformulation of the same problem. Notably, in variants of the problem where one only receives partial samples, our combinatorial algorithm is almost unaffected, whereas the convex approach fails to get off the ground. Ph.D. 2022-08-29T16:27:54Z 2022-08-29T16:27:54Z 2022-05 2022-06-07T15:33:53.299Z Thesis https://hdl.handle.net/1721.1/145023 In Copyright - Educational Use Permitted Copyright MIT http://rightsstatements.org/page/InC-EDU/1.0/ application/pdf Massachusetts Institute of Technology
spellingShingle Hamilton, Linus
Applications and limits of convex optimization
title Applications and limits of convex optimization
title_full Applications and limits of convex optimization
title_fullStr Applications and limits of convex optimization
title_full_unstemmed Applications and limits of convex optimization
title_short Applications and limits of convex optimization
title_sort applications and limits of convex optimization
url https://hdl.handle.net/1721.1/145023
work_keys_str_mv AT hamiltonlinus applicationsandlimitsofconvexoptimization